PhotoGuru went to the "dark side" and "wiser"
I continued to experiment with the prototype of the PhotoGuru program in order to understand whether recognizing the “style” of a photo with something can help the user.
The new version of the updated interface and the program gives only one answer. Most often, this is the name of the photographer and the percentage of “similarity” of the user's photo to the style of one of the six photographers.
If the percentage is below 33%, the system cannot identify the author and concludes that the new “guru” is in front of her and instead of the last name shows the inscription “PHOTO GURU”.
')
I added my photos and portraits of Zhenya Otetskyi to the system and the model began to make mistakes in 31% of cases.
Having edited the samples, I realized that I was starting to adjust the answers to the result. Need another way. It turned out that the system can be taught longer. I increased the number of training cycles to 20 and now the system is almost not mistaken when analyzing photos from a test sample.
In general, the model correctly classifies portraits unknown to her, but the system counted one of my old photos as a photo of Zhenya Otsetstsky and is not quite sure that I myself shot a self-portrait.
But most of all, I am impressed that by training only in portraits, the system was able to confirm with 99% accuracy that Yevgeny Oettsky was shooting landscapes with “trains”.
I tried to find a photo that would not look like any of the authors of the well-known system. I could not get.
The system considered that the photos from Tim Walker are similar to those of Eugene of the Fathers, Mike Brodie shoots like Egor Voinov, the abstraction was made by Maxim Shumilin.
Most likely, the model analyzes photos at the composition level and therefore produces such a result. She cannot assess aesthetics and symbols, and recognizes a white square as a photograph of one of the authors.
Playing with PhotoGuru can be useful for self-learning photos. To do this, you need to increase the number of photographers known to the system, think about protection from "cheaters" and add "awards". For example, the Videnin 99% badge and the big medal badge to the user who collected 99% badges from all photographers.
And if someone receives the status of "PHOTO GURU" - a diploma for uniqueness.
The new version of the updated interface and the program gives only one answer. Most often, this is the name of the photographer and the percentage of “similarity” of the user's photo to the style of one of the six photographers.
If the percentage is below 33%, the system cannot identify the author and concludes that the new “guru” is in front of her and instead of the last name shows the inscription “PHOTO GURU”.
')
if (maxPercent <= 100/numPhotographers*2) { message = " ★ " photoGuru = "Photo Guru" } I added my photos and portraits of Zhenya Otetskyi to the system and the model began to make mistakes in 31% of cases.
Having edited the samples, I realized that I was starting to adjust the answers to the result. Need another way. It turned out that the system can be taught longer. I increased the number of training cycles to 20 and now the system is almost not mistaken when analyzing photos from a test sample.
In general, the model correctly classifies portraits unknown to her, but the system counted one of my old photos as a photo of Zhenya Otsetstsky and is not quite sure that I myself shot a self-portrait.
But most of all, I am impressed that by training only in portraits, the system was able to confirm with 99% accuracy that Yevgeny Oettsky was shooting landscapes with “trains”.
I tried to find a photo that would not look like any of the authors of the well-known system. I could not get.
The system considered that the photos from Tim Walker are similar to those of Eugene of the Fathers, Mike Brodie shoots like Egor Voinov, the abstraction was made by Maxim Shumilin.
Most likely, the model analyzes photos at the composition level and therefore produces such a result. She cannot assess aesthetics and symbols, and recognizes a white square as a photograph of one of the authors.
Playing with PhotoGuru can be useful for self-learning photos. To do this, you need to increase the number of photographers known to the system, think about protection from "cheaters" and add "awards". For example, the Videnin 99% badge and the big medal badge to the user who collected 99% badges from all photographers.
And if someone receives the status of "PHOTO GURU" - a diploma for uniqueness.
Source: https://habr.com/ru/post/446538/
All Articles